IEEE Systems, Man and Cybernetics Magazine - January 2022 - 31

clustering algorithms to predict elevations in ICP [20]. Scalzo
et al. propose an extremely randomized decision-tree
approach to predict hypertension episodes using extracted
morphological waveform features rather than raw, ICP signals
[21]. Fuzzy-based systems have been applied in previous
works for pulse-level analysis and classification of
hypertension [22], [23]. However, they have been applied
primarily on cardiovascular hypertension, not intracranial
hypertension.
We suggest an empirical analysis of the following various
ML algorithms: nearest neighbor, naïve Bayes, Gaussian
process, SVM, QDA, decision tree, AdaBoost, and
MCNNs. We selected ML algorithms that would provide
depth and width to our experiment and are limited to computationally
feasible algorithms that are not computationally
expensive and do not require special computational
requirements. In addition, algorithms of lower complexity
were selected to allow training with limited labeled data.
We also propose the SMOTE-ENN algorithm because the
ICP signals that are unbalanced primarily contain signal
recordings of normal healthy readings.
Methods
An overview of the main data and methods performed in
this article is presented in Figure 1. Two data sources were
prepared to produce two data sets (original and balanced),
which were each used to separately train and test different
models. The performance metrics of each model per data
set were then analyzed to compare the different models.
The following sections discuss the data, methods, and
models in more detail.
Data Source
For this study, we sourced two publicly available ICP data
sets: one each from MIMIC II [24] and CHARIS [25]. The
data sets were cleaned, analyzed, annotated, combined,
and balanced.
MIMIC II
The MIMIC II data set consists of high-resolution signals,
representing the time-series signal of vital signs. It
also contains static clinical records. The purpose of the
data set is to support epidemiologic research and the
assessment of clinical decision-support systems in the
domain of critical-care medicine. The data set was collected
from Beth Israel Deaconess Medical Center (Boston,
Massachusetts, USA). It consists of 25,328 adult
patients' surgical and cardiac records. It was obtained
from 2001 to 2007. The data were entirely deidentified in
accordance to health-act standards. If the same patient
was admitted after 24 h, the data, in that case, were
recorded with a different ID. The time series were
updated at 1 Hz.
All of the data that did not contain ICP signals over
their entire range were discarded. The data set contains
only a small fraction of the data that includes ICP signals.
The segments were cross checked with static patient information
to eliminate duplicate data of the same patient.
Intracranial hypertension is defined as an elevation of
ICP over 20-mm Hg. The data set was imbalanced with
only 10% positive labels (an instance of intracranial hypertension)
among all instances. This imbalance was handled
by the SMOTE-ENN algorithm to create more similar
examples of the positive label.
CHARIS
The CHARIS data set consists of multichannel records of
arterial blood pressure, electrocardiogram, and ICP of
individuals diagnosed with TBI. The purpose of the distribution
of data to researchers is to systematize the analyses
of appropriate physiological signals and construct algorithms
driven by data in search of possible predictors of
critical clinical events for individuals with a significant
brain injury.
1
CHARIS
MIMIC II
2
Data
Cleaning
Preprocessing
Hypertension
Labeling
3
Balanced
Data Set
SMOTE ENN
Models
4
Traditional MLs
SVMs
Tree Based
MCNN
Original
Data Set
5
Model Training
and Testing
6
Performance
Metrics
Model
7
Comparative
Analysis
Figure 1. A diagram illustrating the flow of work. Row 1
shows the source data sets, rows 2 and 3 outline the
processing of the data set, rows 4 and 5 highlight the
ML process, and rows 6 and 7 display the results and
analysis, respectively. SVMs: support vector machines.
January 2022 IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE 31
IEEE Systems, Man and Cybernetics Magazine - January 2022

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - January 2022
